Parallelizing theFuzzy ARTMAP Algorithm on a Beowuif Cluster
Michael Georgiopoulos · 2005
Fuzzy ARTMAP neural networks havebeenproven tobegoodclassifiers ona variety ofclassification problems. However, thetimethatittakesFuzzyARTMAP toconverge toa solution increases rapidly asthenumberofpatterns usedfortraining increases. Inthis paperwe propose a coarse grain parallelization technique, based onapipeline approach, to speed-up FuzzyARTMAP'straining process. Inparticular, we first parallelized FuzzyARTMAP,without thematch-tracking mechanism, andthenweparallelized FuzzyARTMAP withthe match-tracking mechanism. Results runona Beowulf cluster withawellknownlarge database (Forrest Covertype database fromtheUCIrepository) showlinear speedup withrespect to thenumberofprocessors usedinthepipeline.